The Exclusivity Death Spiral

Notes from the Clone Weekend (October 2026)

In one weekend, a four-person team expanded from cloning Adobe to cloning Microsoft Office to cloning AutoCAD. This is not a story about one startup. It is a story about what happens to every business model built on artificial scarcity when AI can make anything and break into anything.

Saturday: Photoshop falls

The weekend started with ArtCraft: seven open-source, clean-room reimplementations of Adobe’s Creative Suite — PhotoCraft, VectorCraft, FilmCraft, LightCraft, and others — built in Rust with Claude doing the heavy lifting. Free, no account, no cloud, no telemetry, no subscription. Some apps early alpha, some already usable.

The stated motive was personal: cancellation fees, rented software, a grudge against the subscription model. The method was the real story: nobody read Adobe’s source code. The AI watched the software behave and rebuilt the behavior from scratch.

Sunday: Office falls, then AutoCAD

Within 24 hours the same team shipped the next wave:

WordCraft — a Word reimplementation that reads and writes DOCX, ~87% of ribbon commands, ~62% overall parity. Independently tested on an Arm Surface and reported as working well.
GridCraft — the Excel clone, ~89% of ribbon commands, ~93% of worksheet functions. That last number is the spicy one: Excel’s calculation engine was supposed to be the moat.
DeckCraft — PowerPoint, ~79% feature coverage, approaching first alpha.
CADCraft — the AutoCAD clone, already a live repo with a parity roadmap. The team’s own estimate: roughly 570 more hours of agent work to broad AutoCAD parity, compressible to under 200 hours elapsed with parallel agents. It was in the Arch Linux package repository within days.

The Discord community is already lobbying for what’s next. The 3D-printing and maker crowd wants the rest of Autodesk. Somebody always wants the next one.

The method is the story

Behind the clones sits a tooling layer that matters more than any single app:

REA (“Reverse Engineer Anything”) — an open project wiring AI agents to binary analysis tools, gaining thousands of stars per day. Instead of asking a model to guess how software works, it investigates: disassembly, call graphs, data structures, behavior traces, evidence-backed reconstruction.
Frontier agents can now do professional reverse engineering. GPT-6 Astra reportedly scores around 88% on binary-analysis benchmarks in a single attempt, near 99% within four. Compiled binaries — the last informational barrier protecting proprietary software — are becoming a speed bump.
Commentators are describing an “open-source multiverse”: every paid launch gets shadowed by a free open copy within days or weeks. Capability diffuses; only distribution and data stay put.

The through-line: anything observable is rebuildable. The only question is who has the angriest users pointing the agents.

A threat model for the next clone

Three factors predict what falls next:

Observable behavior. Can an agent watch the software run and reconstruct what it does? Desktop apps: yes. Web apps: trivially (the browser is an inspection console). Games: yes, and the modding scene verifies results cheaply.
Hate-base energy. Are users motivated enough to aim agents at it? Cancellation fees, per-seat pricing, dongles, forced bundles, shrinking free tiers — resentment is the targeting system.
Moat type. Is the lock-in the code, or something else? This is the sorting hat for everything below.

Tier 1 — already being aimed at

Autodesk (Fusion, Revit, Maya). CADCraft proved the pipeline on 2D drafting. The maker community’s licensing grievances are legendary.
Intuit (QuickBooks, TurboTax). The largest hate-base in consumer SaaS meets fully observable form-filling behavior. The moat is tax-table updates and compliance branding. April comes every year, and so does the rage.
Figma. Browser-observable by definition, beloved-but-resented after the acquisition saga, and canvas/vector editing is well-trodden ground. The hard part isn’t the editor — it’s the multiplayer and ecosystem.

Tier 2 — ripe, needs one motivated sprint

Salesforce / HubSpot. The irony candidate: a CRUD app with a nice UI charging per seat. The software clones easily; the customer data and integrations don’t. Expect the clone to win individuals and small teams while the enterprise stays put.
Notion / Airtable / Monday. Structurally simple (docs plus database plus views), web-observable, and the clone pitch writes itself: local-first, no subscription, your data stays home. The local-first movement (Obsidian, Anytype, AppFlowy) already holds the beachheads.
Music production tools. Passionate users, dongle-and-subscription resentment, deterministic DSP behavior. A free DAW with plugin compatibility is closer than the industry thinks.
Pro video (Resolve Studio, Final Cut). FilmCraft opened this front from the Premiere side. The Studio-tier paywall is next.

Tier 3 — harder, dying differently

Workday / ServiceNow. Clonable as software, nearly unclonable as installed base. These die by agent disintermediation (agents performing the workflow without the UI), not by direct clone.
Bloomberg Terminal. $25K a seat, universally resented, behavior fully observable — but the moat is the live data feed and the chat network. Clone the terminal all day; without the pipe it’s an empty shell.
Epic (health records). The final boss of hate-bases, protected by certification, liability, and hospital IT politics. Unclonable by agents; killable only by legislation.

The sorting hat

If the moat is the software, it gets cloned. If the moat is the data, the network, or the law, it survives.

That one rule predicts every future wave.

The next barrier: agents in apps

The clone wars are about rebuilding apps. The next war is about who gets to operate them. Incumbents who can’t win on software parity will move the moat from the code to the front door:

Agent detection and blocking — bot walls upgraded to “are you an authorized agent?” checks.
API gating — killing unofficial access paths while selling an “official agent tier” at a markup.
UI obfuscation — canvas-rendered interfaces and virtualized layouts that humans use easily but automation can’t scrape.
Legal ToS walls — “automated access prohibited” clauses, enforced selectively. The first “unauthorized agent access” lawsuit will be the test case for whether operating software you pay for through your own agent is a right or a violation.

This will probably fail, for three reasons:

Accessibility law is the skeleton key. Decades of law require apps to be machine-operable (screen readers, voice control, switch devices). An app that blocks agents but permits screen readers is drawing a line the courts will struggle to hold: what is the difference between an agent assisting a disabled user and an agent?
The analog hole. If a human can see it and click it, an agent with a screenshot and a virtual mouse can too. Every block is a speed bump; open-source automation only needs one working path.
Enterprise customers will revolt. The same CIOs paying for AI platforms want agents operating their stack — that productivity story is what they were sold. A vendor blocking agents is blocking the customer’s own return on investment.

The endgame:

Apps that welcome agents become infrastructure. Apps that block agents become islands — and islands get bridged or abandoned.

The winners ship agent-native interfaces (structured APIs, agent protocols, sane agent pricing) and compete on underlying value. The losers spend R&D on detection-and-blocking — a treadmill where blockers pay full cost and the open-source bypass pays marginal cost. We’ve seen this movie. It was called DRM, and the pirates won every round while paying customers suffered.

Reddit, X, and the general law

Apply the same logic to platforms. If Reddit blocks agents, someone builds the agent-readable Reddit and mirrors the data. The data already escaped (every frontier model ingested it), mirroring text-and-votes is a commodity operation, and the contributors — the actual product — are portable. The 2023 API riots already proved the community will migrate when provoked.

X is even more exposed: the exodus already scattered the community across open protocols (AT Protocol, ActivityPub) that are agent-operable by design and cannot be gatekept. Tweets are the most mirrorable content on earth. The wall guards a shrinking treasury.

The general law:

Block the agents → agents build the alternative → the alternative inherits your community → the block becomes the reason you lost.

Twitter’s API pricing minted Mastodon and Bluesky. Reddit’s API pricing minted the Lemmy surge. Each blockade markets its own replacement. And the alternatives don’t even need to win — they just need to exist as credible exit options, because the threat of exit caps the gatekeeper’s pricing power. Same logic as CUDA compatibility versus Nvidia: the alternative disciplines the rent without having to conquer.

The exclusivity death spiral

Every exclusivity business model rests on one hidden assumption: the gated thing is hard to reproduce. AI attacks that assumption from both directions simultaneously:

“AI can make anything” kills supply-side exclusivity. The cost of reproducing any observable artifact collapses toward compute cost.
“AI can break into anything” kills access-side exclusivity. Rate limits, login walls, API gates, DRM — all software barriers, all just puzzles for agents with screenshots and patience.

So what survives when neither supply nor access can be gated? Only things AI can’t reproduce or bypass:

Compute itself. You can’t download a datacenter. Scarcity is physical. (The hyperscalers collect rent from both the incumbents and the rebels.)
Live data feeds. Value is freshness plus network, regenerated every millisecond. A mirror is always stale.
Human trust and relationships. The account team, the implementation partner, the contract with someone to call. Never the CRUD app.
Legal exclusivity. Certification, patents, spectrum. The state enforces what physics won’t. Slowest moat to erode — and it erodes through politics, not technology.
Taste, curation, community. People pay for who’s in the room, not the room.
Outcomes, not access. “Pay when the work is done” instead of “pay monthly for the tool.” When the tool is free, you charge for the result.

The one-line version:

Exclusivity monetizes artificial scarcity. AI manufactures abundance. Every business model built on the former is now shorting the latter — and the margin call is arriving app by app.

Destination certain, timeline humble

High confidence in the destination, humility about the arrival schedule. The layers fall at different speeds:

Months (now): desktop clones, game decompiles, gate bypasses. Observable artifacts fall first because the tooling already exists.
1–3 years: seat-model erosion, workflow disintermediation, the first agent-access lawsuits drawing legal lines.
3–7 years: sieges on data moats, professional-silicon competition, the regulatory-capture endgame. Local hardware either looks prophetic or quaint — no middle ground.
Decade+: legal exclusivity, state-enforced scarcity, everything that moves at the speed of legislation. The last redoubt of rent.

The rational posture is built for destination-certainty with timeline-humility: capability hedges (local machines, open weights, preserved tooling), toll-collector exposure (compute, infrastructure), and no leveraged bets on any single date.